acceptodds
Under review as a conference paper at ICLR 2027

AnyStep-WAM: Budget-Aligned Distillation and Adaptive Inference for World Action Models

Abstract

World-action models (WAMs) couple predictive visual modeling with action generation, but typically rely on iterative denoising with a fixed inference budget, incurring substantial latency while overlooking variation in generation difficulty across interaction states. We introduce AnyStep World Action Model, a general framework for accurate cross-budget prediction and scene-dependent computation allocation. Our budget-aligned teacher-trajectory distillation trains interval-conditioned flow maps using explicit frozen-teacher transitions and shared low-rank adapters, supporting action generation from one-step prediction to multi-step refinement. Building on this capability, a lightweight risk–benefit scheduler predicts teacher-curvature-based difficulty and budget-specific student–teacher fidelity from a single one-step preview, selecting the smallest budget predicted to satisfy risk-adaptive fidelity requirements. We evaluate our framework on three widely used WAMs Motus, FastWAM, and LingBotVA using RoboTwin 2.0. Our method reduces average denoising steps by 60.2%, 49.8%, and 85.28%, respectively, while maintaining baseline task success rates. In particular, our AnyStep training substantially improves model performance under a one-step denoising budget, increasing task success rates by 7.07, 12.08, and 8.94 percentage points on Motus, FastWAM, and LingBotVA, respectively. Experiments on six real-world manipulation tasks further validate its effectiveness.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Robotic Manipulation, Flow Matching, Any-Step Generation, Adaptive Inference, Knowledge Distillation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_WuyWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AnyStep-WAM, a framework for adaptively selecting the computational budget during the deployment of world-action models. The key methodological contributions are distilling a frozen teacher WAM into an interval-conditioned student and training a lightweight model that uses teacher flow path curvature to predict difficulty and required integration fidelity during deployment. The method is evaluated in simulation on RoboTwin 2.0 and in the real world on six bimanual manipulation tasks. In these experiments, inference time decreases by about 40-64% while maintaining success rates comparable to the base policy.

Critical strengths

The paper addresses a highly relevant problem: reducing the high computational burden of world-action models. The proposed method requires extensive additional training, but the experiments show that it can substantially reduce the computational burden at inference time.

Critical weaknesses
  • The paper does not report any information about the statistical significance of the results. All experiments seem to be conducted with a single seed, and the relative difference in success rates among the evaluated methods is mostly within 0 to 2 percentage points. The lack of statistical tests across multiple seeds makes it difficult to assess whether the proposed method, as claimed by the authors, indeed maintains high success rates while reducing computation time.
  • The authors write the following in the motivation of their method: "critical actions require precision, while less sensitive actions allow faster generation with fewer denoising steps". However, the paper is lacking a convincing analysis of whether AnyStep actually allocates more compute to high-precision interactions. At the top of Figure 3, predicted curvature risk is high for inference index 1, although this timestep corresponds to a free-space motion that does not require high precision.
Questions for authors

Clarification requests:

  • Could the authors analyze the statistical significance of their results? For example, statistical tests across multiple training seeds would be highly valuable to assess the impact of AnyStep-WAM on the success rate.
  • Could the authors analyze in what types of situations their method reduces computational effort the most? In Figures 4 and 5, the relationship between task phase/motion and computational budget appears somewhat arbitrary, and it is unclear whether the computational budget is actually highest during high-precision phases.
  • What is the reason for choosing very different percentile values for video (10/20) and action (20/80) similarity? The ablation only evaluates +5/-5 changes to these percentile values, and it would be helpful to analyze the sensitivity with respect to these values more broadly.
  • How were the weighting factors of the four loss terms chosen, and how sensitive is the overall performance of AnyStep-WAM to these values?
  • Could the authors provide details about the training time of their method?

Suggestion for future work:

  • Section 4, in particular Section 4.2, is very dense and sometimes hard to follow. Adding more explanations, especially visualizations, would greatly improve readability here.
  • Adding one more benchmark environment, such as Libero-Plus or Robocasa, would be highly valuable to understand whether the reported behavior also transfers to other robot embodiments, such as single-arm or mobile manipulators.
13 comments
AuthorsOpenReview ↗

Response to W1 and Q1: Statistical significance and robustness across training seeds

Dear Reviewer WuyW,

Thank you for your careful review and constructive feedback. Due to the 5,000-character limit per response, we have divided our replies into multiple comments.

W1 and Q1: Statistical significance and robustness across training seeds

Thank you for highlighting the need for statistical validation. We have added the following results to Appendix H, “Robustness across Training Seeds,” in the revised paper. In addition to the original training seed 42, we trained the student and scheduler with seeds 43–46 for all three backbones. Each evaluation uses 100 trials per task per setting, with SR averaged across tasks and clean/randomized settings. The table reports mean ± sample SD across the five training seeds.

BackboneTraining seedFixed 1-step SR (%)Fixed 2-step SR (%)Fixed 4-step SR (%)Fixed 6-step SR (%)Fixed 8-step SR (%)Fixed 10-step SR (%)Adaptive SR (%)Adaptive mean steps
Motus4282.1286.2187.0486.9487.6488.0987.963.98
Motus4382.2786.1487.1286.9887.7187.9888.114.02
Motus4482.0486.2487.0386.9787.8588.1088.053.95
Motus4582.2586.2286.9286.8587.6188.0287.993.97
Motus4682.3186.1987.0286.9187.7387.9488.023.86
MotusMean ± SD82.20 ± 0.1186.20 ± 0.0487.03 ± 0.0786.93 ± 0.0587.71 ± 0.0988.03 ± 0.0788.03 ± 0.063.96 ± 0.06
FastWAM4289.2889.7690.5691.0692.0091.9491.595.03
FastWAM4389.3289.5790.5491.0991.9791.8291.665.00
FastWAM4489.2489.7790.5590.9792.0691.7891.695.02
FastWAM4589.3889.8290.4391.1092.0891.9591.624.98
FastWAM4689.2089.6090.6591.0392.1991.8791.764.92
FastWAMMean ± SD89.28 ± 0.0789.70 ± 0.1190.55 ± 0.0891.05 ± 0.0592.06 ± 0.0991.87 ± 0.0791.66 ± 0.074.99 ± 0.04
LingBotVA4282.7188.9689.5790.3791.2892.1492.223.68
LingBotVA4382.7388.8689.6590.2791.5092.0192.273.67
LingBotVA4482.8488.8589.4590.4091.3792.0492.213.65
LingBotVA4582.8588.9889.5690.3491.3592.0892.313.57
LingBotVA4682.8389.1089.5590.4191.2592.1592.293.59
LingBotVAMean ± SD82.79 ± 0.0788.95 ± 0.1089.56 ± 0.0790.36 ± 0.0691.35 ± 0.1092.08 ± 0.0692.26 ± 0.043.63 ± 0.05

Across the five training seeds, adaptive inference achieves SRs of 88.03 ± 0.06%, 91.66 ± 0.07%, and 92.26 ± 0.04%, with mean denoising steps of 3.96, 4.99, and 3.63 on Motus, FastWAM, and LingBotVA, respectively. The small SDs indicate consistent performance across training seeds. Adaptive SR remains close to that of the fixed 10-step student while using substantially fewer denoising steps.

AuthorsOpenReview ↗

Response to W2 and Q2: Which phases receive more computation, and the high-risk free-space example

W2 and Q2: Which phases receive more computation, and the high-risk free-space example

Thank you for raising this question. We have added the following content to Appendix E.5, “Task- and Phase-Level Computation Allocation,” in the revised paper. Taking Motus as an example, Table 26 in Appendix K reports results from 100 trials per task per setting. Broad-target transport and placement tasks use fewer denoising steps: Place Empty Cup, Place Container Plate, and Place Burger Fries use only 2.74–2.78 steps on average. Coordinated or geometrically constrained tasks receive larger budgets: Handover Block uses 5.42 steps, Put Object Cabinet uses 5.99, and Hanging Mug uses 6.45, compared with the 50-task average of 3.98.

TaskRepresentative requirementClean mean KKRandomized mean KKAverage KK
Place Empty CupBroad-target transport and placement2.942.542.74
Place Container PlatePlacement into a relatively tolerant target region2.862.702.78
Place Burger FriesShort transport and broad-region placement2.712.802.76
Blocks Ranking RGBColor-based semantic ordering with straightforward pick-and-place motions2.823.112.97
Handover BlockCoordinated dual-arm transfer and release5.874.975.42
Put Object CabinetConstrained placement inside a cabinet6.315.675.99
Hanging MugHandle grasp and constrained hook placement6.466.446.45
All 50 tasks—4.023.943.98

We additionally analyze phase-level statistics from adaptive rollouts of AnyStep-Motus on Hanging Mug, including the risk score and the executed denoising budget. Free-space motion receives the smallest mean budget (4.4 steps), while the grasp/contact transition receives the largest (7.8 steps). The proportion of inference calls using K≥6K\geq6 rises from 34% during free-space motion to 84% during the grasp/contact transition.

Hanging Mug phaseMean risk RRMean KKShare with K≤2K\leq2Share with K≥6K\geq6
Free-space motion0.474.439%34%
Pre-contact alignment0.616.218%63%
Grasp/contact transition0.727.88%84%
Contact manipulation/placement0.677.510%80%

We agree that a single example in Figure 3 does not provide sufficient evidence for the overall relationship between manipulation phases and computation allocation. We apologize for the limited analysis in the original paper and added the task-level and phase-level statistics above to better support and clarify this trend in the revised paper.

AuthorsOpenReview ↗

Response to Q3: Video/action percentile thresholds and wider sensitivity, and Q4: Selection and sensitivity of the loss weights

Q3: Video/action percentile thresholds and wider sensitivity

Thank you for suggesting a broader sensitivity analysis. We have added the following results to Appendix E.6, “Video/Action Percentile Thresholds and Sensitivity,” in the revised paper. We compute the easy and hard similarity thresholds from the specified percentiles of each modality’s offline training-set distribution. Predicted risk interpolates between these two threshold values, and the scheduler selects the smallest budget satisfying both video and action fidelity requirements. If no budget qualifies, it selects the budget with the largest summed predicted fidelity.

The asymmetric defaults reflect the different roles of the two modalities. Video similarity captures agreement in predicted visual dynamics, including appearance and background changes that may be less relevant to control precision. Its 10th/20th percentile thresholds therefore provide a relatively permissive scene-consistency check. Action similarity measures agreement in the commands executed by the robot. The 20th/80th percentile thresholds allow a permissive requirement in easy states and require higher action fidelity in high-risk states.

We have expanded the Motus sensitivity evaluation as shown below, keeping the student and scheduler fixed and changing only the inference thresholds. SR and mean steps are averaged across clean and randomized settings.

Configuration (Motus)Video percentiles (easy/hard)Action percentiles (easy/hard)SR (%)Mean steps
Default10 / 2020 / 8087.963.98
Joint −55 / 1515 / 7587.603.08
Joint +515 / 2525 / 8587.935.85
Video only: lower5 / 1020 / 8087.813.54
Video only: higher20 / 4020 / 8087.864.67
Action only: lower10 / 2010 / 4087.022.61
Action only: lower hard threshold10 / 2020 / 5087.473.25
Action only: higher10 / 2030 / 9087.895.16
Shared 10 / 2010 / 2010 / 2086.742.23
Shared 20 / 8020 / 8020 / 8087.626.71

SR is relatively stable over most of the tested range: 8 of the 10 configurations remain within 0.49 percentage points of the default, while mean steps vary from 2.23 to 6.71. With action percentiles fixed at 20/80, varying the video percentiles changes SR by at most 0.34 points. More permissive action settings of 10/40 and 10/20 reduce SR by 0.94 and 1.22 points, respectively. Higher thresholds increase computation without improving SR over the default. These results support the default as a practical SR–computation trade-off, while showing that SR is less sensitive than the computational budget across most tested settings.

Q4: Selection and sensitivity of the loss weights

Thank you for asking about the loss weights. We have added the following results in Appendix C.2 of the revised paper. We use (λFM,λend,λint,λrec)=(0.20,0.50,0.30,0.20)(\lambda_{\mathrm{FM}},\lambda_{\mathrm{end}},\lambda_{\mathrm{int}},\lambda_{\mathrm{rec}})=(0.20,0.50,0.30,0.20) based on the roles of the four terms. The endpoint loss receives the largest weight because the one-step endpoint is both an important inference mode and the preview reused by larger budgets. The interval loss provides direct supervision for budget-aligned finite-time transitions. The flow-matching term maintains instantaneous-velocity prediction, while the recovery term supervises continuation from preview-recycled student states. These weights were fixed across Motus, FastWAM, and LingBotVA and were not tuned on the test set.

To evaluate sensitivity on Motus, we test two alternative combinations. Alternative A assigns more weight to endpoint and interval supervision, whereas Alternative B assigns more weight to flow matching and recovery. For each setting, we retrain both the AnyStep student and its scheduler under the same protocol.

Configuration (Motus)(λFM,λend,λint,λrec)(\lambda_{\mathrm{FM}},\lambda_{\mathrm{end}},\lambda_{\mathrm{int}},\lambda_{\mathrm{rec}})Fixed 1-step SR (%)Fixed 4-step SR (%)Fixed 10-step SR (%)Adaptive SR (%)Adaptive mean steps
Default(0.20, 0.50, 0.30, 0.20)82.1287.0488.0987.963.98
Alternative A(0.10, 0.60, 0.40, 0.10)82.3887.1587.9287.783.86
Alternative B(0.30, 0.40, 0.20, 0.30)81.7686.8188.1687.844.19

Performance remains stable across the tested configurations: fixed-budget SR differs from the default by at most 0.36 percentage points, and adaptive SR by at most 0.18 points, with mean budgets of 3.86–4.19 steps. Alternative A has slightly higher one- and four-step SR and a lower mean adaptive budget, while Alternative B has slightly higher ten-step SR and a higher mean adaptive budget. The default achieves the highest observed adaptive SR among the three configurations. These results suggest limited sensitivity to the two tested weight redistributions.

Reviewer_5WzaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

(1) This paper accelerates World Action Models (WAMs) by enabling adaptive, any-step inference within a single model, rather than a fixed denoising budget.

(2) It proposes budget-aligned integral flow-map distillation: a pretrained WAM is adapted to predict transitions over arbitrary time intervals, supervised by finite displacements along frozen-teacher trajectories rather than derivative-based targets.

(3) A lightweight predictor uses one-step-preview signals and learnable queries to predict a risk score and per-budget benefit scores, which drive selection of the smallest sufficient budget; the preview is recycled into execution, requiring no online teacher at inference.

(4) On RoboTwin 2.0 with three WAM backbones, the method preserves success rates near full-budget baselines with far fewer steps, and real-robot experiments confirm its practicality.

Critical strengths

S1: Practical acceleration of mainstream WAMs via adaptive any-step inference. The method accelerates current mainstream WAMs (Motus, FastWAM, LingBotVA), and the any-step inference adaptively decides when more denoising steps are needed, balancing inference speed against action precision. On RoboTwin 2.0, adaptive inference stays within 0.24 percentage points of the full-budget baselines while reducing mean denoising steps from a fixed 10 (or 25/50 for the serial backbone) to 3.98/5.03/3.68.

S2: Consistently outperforms a direct transfer of AnyFlow. On RoboTwin 2.0, the method outperforms AnyFlow at every fixed denoising budget (e.g., average SR of 82.12/89.28/82.71 vs. 13.75/32.01/49.96 at one step, on Motus/FastWAM/LingBotVA, respectively), and improves one-step SR over the base models by 7.07/12.08/8.94 points.

Critical weaknesses

W1: Training cost is unreported and potentially heavy relative to alternative methods.

The paper does not clarify whether its two-stage training — budget-aligned distillation followed by scheduler training — incurs substantially higher cost than the methods it compares against, despite the method requiring extensive teacher and student rollouts.

According to the method description:

  • each distillation sample requires a fine-grained frozen-teacher reference trajectory (solver step count N never disclosed) plus teacher restart rollouts for the recovery loss;
  • each scheduler-training sample requires rehearsing the preview-recycled execution path for every candidate budget K ∈ {1,…,10} — by my estimate, roughly 45 student evaluations and about ten teacher rollouts per sample.

The paper reports no per-sample teacher/student NFE breakdown, no total training time, no training-cost comparison with the baselines (e.g., AnyFlow or standard fine-tuning), and no statement on whether reference trajectories are cached or recomputed.

Questions for authors

Q1: Missing cheap-distillation baselines at moderate fixed budgets.

Table 6 shows that the un-distilled base models degrade at the step counts used by AnyStep's adaptive inference, and Flash-WAM is compared only under a one-step budget. However, the paper does not include simpler few-step distillation baselines applied to the same backbones — for example, consistency-style distillation on top of the flow-matching objective, or DMD — evaluated at a fixed moderate budget (e.g., 4–5 steps). Such baselines require no adaptive scheduling and would likely involve much lower training cost than the proposed two-stage pipeline, which relies on dense teacher rollouts, per-budget teacher restarts, and offline scheduler labeling. Have the authors considered this comparison? If a simply distilled model at a fixed 4–5 step budget approached the adaptive results (e.g., 87.96 at a mean of 3.98 steps on Motus, 91.59 at 5.03 on FastWAM), the practical advantage of the framework relative to its training cost would be considerably less clear.

Q2: Comparison with recent few-step/one-step acceleration methods.

Beyond AnyFlow and Flash-WAM, a growing body of recent work studies few-step or one-step inference for flow-based VLA and WAM policies. Could the authors compare AnyStep against representative methods from this line of work, or at least clarify its advantages over them in terms of effectiveness and training cost? Without such a comparison, it is difficult to judge the relative merit of the proposed framework among existing acceleration techniques.

8 comments
AuthorsOpenReview ↗

Response to W1: Training cost and offline computation

Dear Reviewer 5Wza,

Thank you for your careful review and constructive feedback. Due to the 5,000-character limit per response, we have divided our replies into multiple comments.

W1: Training cost and offline computation

1.Training cost:

Thank you for highlighting the need to report training cost. We have added the following content in Appendix C.7 and C.8 of the revised paper. The table compares the GPU-hours for the AnyFlow and AnyStep training stages, starting from a given Base checkpoint. AnyStep training includes student distillation with teacher trajectories and recovery supervision. The corresponding AnyStep/AnyFlow ratios are 1.70×, 1.43×, and 1.34× for Motus, FastWAM, and LingBotVA, respectively. Our primary goal is to reduce the inference latency of existing WAMs through AnyStep training, which enables performance improvement across denoising budgets. We validate this capability across joint video–action, action-only, and sequential video-to-action architectures. Thank you for highlighting offline cost and scaling, and we will investigate improvements in training efficiency in future work.

BackboneAnyFlow GPU-hoursAnyStep GPU-hoursAnyStep / AnyFlow
Motus75.1127.61.70×
FastWAM130.1186.01.43×
LingBotVA424.6568.91.34×

During distillation, the original fine-grained teacher reference trajectories are cached and reused for a fixed condition and initial noise. The teacher reference solver uses each Base WAM’s original inference denoising budget NN (e.g., N=10N=10 for Motus). Budget-aligned interval supervision samples a budget KK and an interval, then retrieves the corresponding teacher states from the cache to construct the target. For recovery supervision, the frozen teacher is restarted once from the first recycled student state for each evaluated budget K>1K>1, and the resulting continuation provides references for all remaining intervals. The distillation GPU-hours reported in the table above include reference-trajectory generation and teacher continuations for recovery supervision.

Figure 4 shows that teacher-integrated targets produce more concentrated target norms and more stable gradients than finite-difference supervision in our experiments. Under finite differences, 45.9% of training steps exceed the gradient-clipping threshold of 1.0. Table 18 reports the following SRs:

BackboneMethod1-step SR (%)4-step SR (%)10-step SR (%)
MotusAnyFlow13.7581.5787.38
MotusAnyStep82.1287.0488.09
FastWAMAnyFlow32.0142.5577.01
FastWAMAnyStep89.2890.5691.94
LingBotVAAnyFlow49.9684.2889.70
LingBotVAAnyStep82.7189.5792.14

These results show higher SR for our trajectory-based supervision on the evaluated WAMs, particularly at small budgets, at the additional offline cost quantified above.

2. Offline computation: We added this analysis to Appendices C.5, C.6, and C.9 of the revised paper. The scheduler is designed primarily for lightweight, fast online budget selection. For scheduler training, we reuse the existing demonstrations without collecting additional environment rollouts. After freezing the student, we generate and save preview features, risk labels, and fidelity labels once for reuse throughout scheduler optimization. The offline teacher/student denoising evaluations are included in the labeling cost. For a fixed condition and initial noise with one shared preview, the student label-generation count is

NS,label=1+∑K=110(K−1)=46.N_{S,\mathrm{label}}=1+\sum_{K=1}^{10}(K-1)=46.

This comprises one shared preview and 45 continuation evaluations. The teacher produces one original reference trajectory and nine budget-specific continuations, with total forward evaluations

NT,label=NT,0+∑K=210NT,rec(K).N_{T,\mathrm{label}}=N_{T,0}+\sum_{K=2}^{10}N_{T,\mathrm{rec}}(K).

Each continuation starts from the first recycled state for its budget and supplies references for all remaining intervals. Once features and labels are saved, scheduler optimization requires no additional teacher or student evaluations.

BackboneLabel generation GPU-hoursScheduler optimization GPU-hours
Motus53.96.2
FastWAM37.25.9
LingBotVA77.86.3

Scheduler optimization uses four H100 GPUs for 4,000 iterations with a global batch size of 256. Its 5.9–6.3 GPU-hours correspond to approximately 1.5–1.6 hours on four GPUs. The separate one-time labeling cost is 37.2–77.8 GPU-hours. We appreciate the reviewer’s attention to offline cost, and will investigate lower-cost scheduler training and label generation in future work.

AuthorsOpenReview ↗

Response to Q1: Comparisons with distillation baselines

Q1: Comparisons with distillation baselines

1. Moderate-budget comparisons with AnyFlow. Thank you for highlighting this point. Due to space constraints, Figure 4(c) presents SR at 1, 2, and 10 steps, while Appendix E, Table 6 of the original submission reports the complete fixed-budget results, including 4 and 6 steps. These results now appear in Appendix F.1, “Additional Results on RoboTwin 2.0 across Fixed Denoising Budgets,” Table 18 of the revised paper.

BackboneAnyFlow, 4-step SR (%)AnyStep + adaptive SR (%)Adaptive mean steps
Motus81.5787.963.98
FastWAM42.5591.595.03
LingBotVA84.2892.223.68

Compared with four-step AnyFlow, our full adaptive framework improves observed SR by 6.39, 49.04, and 7.94 percentage points on Motus, FastWAM, and LingBotVA, respectively, using 3.98, 5.03, and 3.68 mean steps. These results support the effectiveness of the full framework under the reported inference budgets.

At the time of our original submission, AnyFlow [1] and Flash-WAM [2] were the latest relevant acceleration methods we had identified, and we included both as baselines. Flash-WAM targets one-step generation per modality, so Table 1 evaluates it at this intended budget.

2. LCM and DMD2 comparisons. We have added the following content to Appendix F.2, “LCM and DMD2 Comparisons,” in the revised paper. We adapted Latent Consistency Distillation [3] and DMD2 [4] to Motus and evaluated them at fixed budgets of 3, 4, 5, and 6 steps. We also added AnyFlow results at 3 and 5 steps. All methods start from the same Motus Base checkpoint and use the same training data and evaluation protocol. SR is averaged across tasks and clean/randomized settings, following the main paper.

MetricAnyStep (Ours)AnyFlowAnyFlowAnyFlowAnyFlowLCMLCMLCMLCMDMD2DMD2DMD2DMD2
SR (%)87.9670.8081.5783.5084.8580.4282.5284.3284.8981.7483.6884.7185.31
Denoising steps3.98 (mean)345634563456

Consistency distillation[3] encourages endpoint agreement along teacher-connected noise levels, but does not explicitly supervise the finite-interval velocity maps traversed by each inference budget. Moreover, as analyzed by FlashWAM [2], the standard LCM parameterization weakens the effective learning signal for precision-critical actions in the low-noise regime. DMD2 [4] partially alleviates this limitation through distribution matching, and therefore improves over LCM at the same budgets. However, its objective aligns the student’s output distribution with the teacher distribution rather than enforcing teacher-aligned transitions for each observation, noise realization, and budget. Its dynamically trained fake-score critic must also track a non-stationary student distribution, which is challenging under the dimensional and noise-distribution asymmetry between video and action.

In contrast, AnyStep directly supervises budget-aligned finite-interval transitions, dedicated endpoint prediction, and recovery from preview-recycled student states. It additionally selects the budget according to state-dependent risk and budget-specific fidelity. AnyStep achieves 87.96% SR at 3.98 mean steps, exceeding four-step LCM and DMD2 by 5.44 and 4.28 percentage points, respectively. Even at six steps, LCM and DMD2 reach only 84.89% and 85.31%, remaining 3.07 and 2.65 points below AnyStep while using approximately 51% more denoising steps.

AuthorsOpenReview ↗

Response to Q2: Comparisons of recent methods

Q2: Comparisons of recent methods

Thank you very much for pointing this out. At submission, AnyFlow [1] and Flash-WAM [2] were the latest relevant acceleration methods we had identified and included as baselines. RealtimeWAM [6], 1Step-WAM [7], and OSWAM [5] appeared after the submission deadline. We added references to these concurrent works [5-7] , and discuss these methods in the revised paper (Section 1). Our motivation is to make a single student usable across denoising budgets for existing WAM architectures and to select an appropriate budget for each inference context. This requires both accurate prediction across budgets and a mechanism for allocating computation. The methods below address related but different problems.

MethodMotivation and main approachDifference from our research question
AnyFlow [1]Preserve video-generation quality as the sampling budget changes, using on-policy flow-map distillation.Shares the goal of learning across budgets, but does not address the precision requirements of action generation or context-dependent budget selection for robotic control.
Flash-WAM [2]Prevent action-quality degradation during joint video–action distillation by accounting for modality-specific noise regimes and gradient scaling.Prioritizes low-step generation; it does not explicitly optimize performance as the denoising budget increases or select a budget for each context.
OSWAM [5]Make one-step action generation accurate by replacing unstructured noise initialization with imagination-token priors and distilling privileged future information.Improves the starting point for one-step prediction; it does not study training and selecting among multiple denoising budgets.
RealtimeWAM [6]Reduce action-denoising iterations and waiting between experts through teacher-anchored consistency distillation and block-level execution overlap.Optimizes one-step action quality and execution parallelism; it does not study training across multiple denoising budgets or context-dependent budget selection.
1Step-WAM [7]Approximate multi-step actions with a lightweight first-step velocity calibrator, bandwidth-constrained decoding, and overlapping video/action execution.Targets an accurate one-step approximation with limited trainable parameters; it does not learn budget-dependent transitions and online budget selection.
AnyStep (Ours)Train budget-aligned transitions with teacher-trajectory and recovery supervision, then select a budget using preview-based risk and budget-specific fidelity.Jointly addresses prediction across 1–10 steps and context-dependent allocation, across joint video–action, action-only, and sequential video-to-action WAMs.

The distinction is therefore the research question, not simply the number of steps achieved. AnyFlow overlaps with our flexible-budget training goal, while the other methods emphasize low-step prediction quality and execution efficiency. None of these methods studies the combined problem of budget-flexible WAM training and context-dependent denoising-budget allocation addressed here. Our fixed-budget experiments evaluate the training component, and the scheduler comparisons evaluate the additional value of selecting budgets for individual contexts.

[1] Gu, Yuchao, et al. “Anyflow: Any-step video diffusion model with on-policy flow map distillation.” European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

[2] Akbari, Arman, et al. “Flash-WAM: Modality-Aware Distillation for World Action Models.” arXiv preprint arXiv:2606.05254 (2026).

[3] Luo, Simian, et al. “Latent consistency models: Synthesizing high-resolution images with few-step inference.” arXiv preprint arXiv:2310.04378 (2023).

[4] Yin, Tianwei, et al. “Improved distribution matching distillation for fast image synthesis.” Advances in neural information processing systems 37 (2024): 47455-47487.

[5] Zhang, Guowen, et al. “OSWAM: An Effective World Action Model with One-Step Action Generation.” arXiv preprint arXiv:2610.35850 (2026).

[6] Lv, Chengtao, et al. “RealtimeWAM: One-Step Asynchronous World Action Models.” arXiv preprint arXiv:2610.06617 (2026).

[7] Su, Rundong, et al. “1Step-WAM: Your World Action Model but Faster using 1-Step Action Generation.” arXiv preprint arXiv:2610.15625 (2026).

Reviewer_EZczOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper speeds up inference of world action models (WAMs) in two ways. (1) Budget-aligned distillation: LoRA adapters turn a frozen WAM into a flow map that can run at any budget K from 1 to 10. Each interval jump is supervised by the corresponding displacement along a frozen teacher’s ODE trajectory, plus endpoint and recovery terms. (2) A risk-benefit scheduler that predicts, from a single one-step preview, a “risk” (teacher trajectory curvature) and a per-budget “benefit” (student–teacher velocity agreement). It then picks the smallest K whose predicted benefit exceeds a risk-dependent threshold. On RoboTwin 2.0, across Motus, FastWAM and LingBotVA, the method uses about 4–5 denoising steps on average while keeping success rate (SR) close to the base models. It also improves one-step SR by 7–12 pt. Six real-robot tasks are included.

Critical strengths

1.The distillation works consistently across three different WAM architectures. 2.The writing is clear and well-expressed.

Critical weaknesses

1.The scheduler’s contribution is not clearly demonstrated. The headline efficiency results compare adaptive inference with the undistilled base model at its full budget, so they mix the gains from distillation with those from adaptive allocation. This paper does not systematically compare adaptive inference with the distilled student at a fixed step count of matching mean budget, in either success rate or inference speed. Table 6 suggests that a fixed budget of a few steps already performs close to the adaptive results. It is therefore unclear how much of the claimed speedup comes from the scheduler itself rather than from distillation. 2.The “risk” signal does not match its motivation. The paper motivates risk as task sensitivity (“critical actions require precision”). However, Sec. 4.2 defines it as teacher denoising curvature and states that it is not a failure probability. Curvature reflects how multimodal the posterior is, not how sensitive the task is. For example, a precise but consistently demonstrated grasp can have low curvature. The paper lacks evidence linking risk to task outcomes.

Questions for authors

1.Could the authors compare adaptive inference with the distilled student run at a fixed step count matching the adaptive mean budget, reporting both success rate and inference latency? This would isolate the scheduler’s contribution from that of distillation. 2.Can the authors provide enough evidence that the risk signal is associated with task difficulty or precision requirements? For example, how does risk vary across tasks of different difficulty and across manipulation phases (e.g., free-space motion vs. contact-rich interaction), and does it correlate with failures?

7 comments
AuthorsOpenReview ↗

Response to W1 Q1: Scheduler gains beyond distillation

Dear Reviewer EZcz,

Thank you for your careful review and constructive feedback. Due to the 5,000-character limit per response, we have divided our replies into multiple comments.

W1 Q1: Scheduler gains beyond distillation

Thank you for this suggesttion. We added these results to Section 5.1 and Appendix I in the revised paper. Using the same students, our scheduler improves SR over AnyStep+Gate by 0.70, 1.43, and 1.88 points while reducing mean steps by 21.8%, 24.5%, and 23.5% on Motus, FastWAM, and LingBotVA, respectively (Table 1). Compared with fixed ten-step students, mean steps decrease by 60.2%, 49.8%, and 63.2%, with SR changes of −0.13, −0.35, and +0.08 points, respectively (Tables 1 and 18).

1. To separate the contributions of distillation and scheduling, we extend Appendix F.1, Table 18 with Base and AnyStep results at neighboring integer budgets and a fixed 10-step reference budget, reporting actual latency on a single NVIDIA A100 (Appendix I, Table 24). We then compare adaptive inference with the same distilled student at nearby fixed budgets to isolate the scheduler’s additional contribution. We included these comparisons in the revised paper and distinguish the full framework’s efficiency gains from the additional gains of adaptive scheduling.

Compared with the same student at the nearest fixed budgets (4, 5, and 4 steps for Motus, FastWAM, and LingBotVA), adaptive inference improves observed SR by 0.92, 0.61, and 2.65 percentage points, respectively. For Motus and LingBotVA, latency decreases 11.8% and 7.1%. Compared with the six-step FastWAM student, adaptive inference achieves 0.53 points higher SR with 9.8% lower latency. Compared with the fixed 10-step students, adaptive SR remains within 0.35 percentage points while latency decreases by 41.4%–56.2%. At the neighboring fixed budgets, AnyStep improves SR over Base by 3.14–3.76 percentage points.

BackboneStepsSR (%)Latency (ms/call)
Motus (base)383.42765
Motus (ours fixed budget)386.73793
Motus (ours adaptive budget)3.98 (Adaptive)87.96859
Motus (base)483.90969
Motus (ours fixed budget)487.04974
Motus (base)1087.841932
Motus (ours fixed budget)1088.091959
FastWAM (base)587.22271
FastWAM (ours fixed budget)590.98276
FastWAM (ours adaptive budget)5.03 (Adaptive)91.59295
FastWAM (base)687.74309
FastWAM (ours fixed budget)691.06327
FastWAM (base)1091.83496
FastWAM (ours fixed budget)1091.94503
LingBotVA (base)385.582914
LingBotVA (ours fixed budget)388.722986
LingBotVA (ours adaptive budget)3.68 (Adaptive)92.223247
LingBotVA (base)486.153403
LingBotVA (ours fixed budget)489.573495
LingBotVA (base)1090.395972
LingBotVA (ours fixed budget)1092.146033

2. Analysis of incremental gains from adaptive scheduling. We further analyze Motus on all five tasks where the original Base Motus achieves SR below 70%, averaged across clean and randomized settings. Adaptive inference achieves 62.40% SR at 4.67 mean steps, improving over the fixed 4-step and 10-step AnyStep students by 7.00 and 5.40 percentage points, respectively. Holding the distilled student fixed allows us to assess the adaptive scheduler’s additional contribution separately from distillation, including a comparison at comparable budgets. Many RoboTwin tasks already achieve relatively high SR with both Base and AnyStep, so averaging across all tasks dilutes the larger gains observed on difficult manipulation tasks. We thank the reviewer for encouraging this analysis and included these task-level comparisons in the revised paper (Table 2).

TaskBase 2 StepAnyStep 2 StepBase 3 StepAnyStep 3 StepBase 4 StepAnyStep 4 StepBase 6 StepAnyStep 6 StepBase 10 StepAnyStep 10 StepAdaptive SR (%)Adaptive Steps
Blocks Ranking Size57.0057.5056.0057.5057.5058.0059.0059.5069.0066.0072.503.81
Hanging Mug24.0034.0024.5034.0025.0039.0030.5034.5038.0036.0044.506.45
Move Can Pot44.0045.5045.0053.0045.5053.5049.5054.5054.0050.5060.003.46
Place Mouse Pad52.5056.5052.5057.0053.0057.5059.5060.0067.0066.5066.503.84
Scan Object56.5065.0058.0064.0058.5069.0063.5067.5066.5066.0068.505.81
Average46.8051.7047.2053.1047.9055.4052.4055.2058.9057.0062.40(+7.00 pp vs AnyStep fixed 4 step, + 5.40 pp vs AnyStep fixed 10 step)4.67
AuthorsOpenReview ↗

Response to W2 Q2: Risk interpretation and validation

W2 Q2: Risk interpretation and validation

Thank you for raising this important distinction. We have added the following results to Appendix E.2 and Appendix E.4, “Risk across Tasks and Manipulation Phases,” in the revised paper. To avoid target leakage, no success or failure labels are used to train the scheduler. We have revised the terminology to clarify that our risk score measures teacher-trajectory variation as a proxy for conditional denoising difficulty, rather than a calibrated failure probability. Higher curvature indicates a more rapidly varying transport trajectory, and our additional analysis shows that it is positively associated with more demanding manipulation phases and higher task-level failure rates. Nevertheless, physical precision and trajectory variation do not always coincide; therefore, risk is not used alone, but jointly with budget-specific teacher–student fidelity.

1. More Ablation: We first compare the complete scheduler against three controlled variants:

Scheduler variantRiskFidelityRisk-conditioned thresholdClean SR (%)Randomized SR (%)Avg. (K)
Risk-only✓––84.783.93.97
Benefit-only–✓–86.385.63.95
Constant threshold✓*✓–86.986.44.01
Full scheduler✓✓✓88.1287.803.98

Risk-conditioned thresholds improve clean and randomized SR by 1.22 and 1.40 percentage points over constant thresholds, respectively, at comparable mean budgets (3.98 vs. 4.01 steps). Risk-only performs worse at a nearly identical mean budget, supporting the contribution of budget-specific fidelity prediction.

2. The relationship between risk and task difficulty and precision requirements: All additional analyses below use fixed-4-step trajectories from the distilled AnyStep-Motus model to examine the relationship between risk, task requirements, manipulation phases, and task-level failure rates.

TaskIndependently annotated manipulation requirementEpisodesMean (C_v) (95% CI)Mean (C_a) (95% CI)Fixed-4 SR (%)
Place Empty CupTolerant transport and placement into an open target region200.18 [0.16, 0.20]0.41 [0.38, 0.44]100
Handover BlockCoordinated dual-arm grasp, transfer, and release200.26 [0.23, 0.29]0.62 [0.59, 0.65]80
Blocks Ranking RGBVisually conditioned repeated grasping and ordered placement200.45 [0.42, 0.48]0.45 [0.42, 0.48]95
Hanging MugHandle grasp and constrained hook placement with tight pose tolerance200.55 [0.52, 0.58]0.56 [0.53, 0.59]35

The four tasks in the table illustrate how teacher-trajectory difficulty varies across manipulation requirements and observed task success rates.

Within tasks, risk also varies systematically across manipulation phases:

TaskEpisodesFree-space motionPre-contact alignmentGrasp/contact transitionContact manipulation/placementContact minus free-space (R)
Hanging Mug200.430.580.730.69+0.28
Blocks Ranking RGB200.360.490.650.60+0.27

Note: Contact minus free-space is computed as the equally weighted mean risk of the grasp/contact transition and contact manipulation/placement phases, minus the mean risk during free-space motion.

Risk increases during alignment and contact-rich stages relative to free-space motion. Together with the scheduler ablation, these results support our intended interpretation: curvature is an informative proxy for denoising difficulty, while budget-specific fidelity determines whether additional denoising is actually beneficial. This motivates combining both signals.

Dear Reviewer EZcz,

Thank you again for your careful review and constructive feedback. We have provided point-by-point responses to your comments in our replies above.

With two days remaining in the discussion phase, we would be grateful to know whether our responses and additional evidence have adequately addressed your concerns. If you find that our rebuttal and additional evidence have adequately addressed them, we would be deeply grateful for your consideration of revisiting the evaluation. If any questions remain or further clarification would be helpful, we would be happy to continue the discussion.

We look forward to hearing from you. Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_dhCGOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary
  1. The paper proposes AnyStep-WAM, which supports multiple denoising budgets by learning finite-interval transitions from frozen teacher trajectories.
  2. A scheduler predicts denoising difficulty and budget-specific teacher–student fidelity from a one-step preview to select the step count. The preview is reused to reduce additional denoising computation.
  3. The method is evaluated on three WAM backbones, 50 RoboTwin tasks, and six real-world robotic tasks. The paper reports reduced NFE and inference latency while maintaining mean success rates comparable to the original full-budget models.
Critical strengths
  1. Efficiency improvements are demonstrated across multiple models and real-world tasks. The method is applied to three backbones and achieves reported per-call speedups of 1.7~6.1x on the real robot. Measuring actual latency alongside NFE reductions supports its practical value.

  2. Distillation improves performance at low NFE. Under the same one-step budget, the method outperforms Flash-WAM by 1.3~7.5 percentage points in success rate. The training stability of teacher-based targets and the effects of key distillation losses provide supporting evidence for the proposed approach to low-budget inference.

Critical weaknesses
  1. [Major] Uncertainty remains about the incremental gains from adaptive scheduling. At comparable step budgets, the improvement over fixed four-step AnyStep varies across backbones: 0.92 percentage points on Motus and 2.65 points on LingBotVA (Tables 1 and 6). Results across independent training seeds or confidence intervals are not reported, making it difficult to assess the robustness of these gains and their value relative to the added complexity.

  2. [Major] Comparisons with adaptive inference baselines are incomplete. The fixed-NFE experiments demonstrate the benefits of distillation, but there is no adaptive-step baseline tailored to the original, undistilled Base model. Comparisons at matched actual latency are also lacking, making it difficult to establish the additional benefit of introducing both the proposed training procedure and scheduler.

  3. The Gate baseline uses a different execution procedure. When the selected step count exceeds one, AnyStep+Gate reruns the complete trajectory after the preview, whereas the proposed method reuses the preview (Appendix D). The comparison therefore reflects differences in both scheduling and execution. A comparison with the same preview-reuse procedure is needed to isolate the scheduler’s contribution.

  4. Offline computational costs and their scaling are insufficiently characterized. GPU counts and training iterations are provided, but total GPU-hours and memory costs, including teacher trajectory generation and budget-specific labeling, are not reported. How these costs grow with model size and the number of candidate budgets is also unclear, limiting an assessment of practicality that accounts for the additional training cost.

  5. Quantitative validation of the step-selection signals is limited. The authors explicitly describe curvature-based difficulty and teacher–student fidelity as proxies rather than direct measures of task success. However, there is limited quantitative evidence showing how well these signals identify situations that require additional steps or predict the actual benefit of additional computation.

  6. The independent contributions of the scheduler components are unclear. The paper includes interval, endpoint, and recovery loss ablations, threshold sensitivity experiments, and clean-only scheduler training. However, risk-only, benefit-only, and risk-independent constant-threshold comparisons are missing, leaving the need to combine both signals with risk-conditioned selection insufficiently established. An ablation of the standard flow-matching loss is also not reported.

  7. The presentation makes the central contribution difficult to identify. Multiple losses, proxies, and execution procedures are introduced, but the connections between the problem addressed by each component and its experimental evidence are not sufficiently clear. This makes it difficult to distinguish the central ideas from auxiliary design choices.

Questions for authors
  1. [Additional validation] Beyond the fixed-budget results in Table 6, does adaptive scheduling retain its advantage at matched actual latency? How does it compare with an adaptive-step baseline tailored to the original Base model? Can results across independent training seeds or confidence intervals be provided?

  2. [Additional validation] How well do the predicted difficulty and fidelity predict task-performance improvements from additional steps? Can comparisons with risk-only, benefit-only, and risk-independent constant-threshold variants establish the need for each component?

  3. [Clarification] Do the step counts reported for AnyStep+Gate in Table 1 include the preview cost? Does the proposed scheduler retain its advantage when both methods use the same preview-reuse procedure?

  4. [Clarification] What are the total GPU-hours and memory costs, including teacher trajectory generation and budget-specific labeling? How do these costs grow with model size and the number of candidate budgets?

9 comments
AuthorsOpenReview ↗

Response to W1: Robustness Across Training Seeds

Dear Reviewer dhCG,

Thank you for your careful review and constructive feedback. Due to the 5,000-character limit per response, we have divided our replies into multiple comments.

W1: Training seeds

Thank you for highlighting the need for validation across training seeds. We have added the following results across five different training seeds to Appendix H and Section 5.1 of the revised paper. We address the concern about “incremental gains from adaptive scheduling” in our response to W2 and Q1.

All training in the original submission used seed 42. We additionally trained the student and scheduler with seeds 43, 44, 45, 46 for all three backbones. The table reports fixed-budget SR, alongside adaptive SR and mean steps. Each evaluation uses 100 trials per task per setting, with SR averaged across tasks and clean/randomized settings. We report mean ± sample SD across the five training seeds.

Across the five training seeds, adaptive inference achieves SRs of 88.03 ± 0.06%, 91.66 ± 0.07%, and 92.26 ± 0.04%, with mean denoising steps of 3.96, 4.99, and 3.63 on Motus, FastWAM, and LingBotVA, respectively. The small SDs indicate consistent performance across training seeds. Adaptive SR remains close to that of the fixed 10-step student while using substantially fewer denoising steps.

BackboneTraining seedFixed 1-step SR (%)Fixed 2-step SR (%)Fixed 4-step SR (%)Fixed 6-step SR (%)Fixed 8-step SR (%)Fixed 10-step SR (%)Adaptive SR (%)Adaptive mean steps
Motus4282.1286.2187.0486.9487.6488.0987.963.98
Motus4382.2786.1487.1286.9887.7187.9888.114.02
Motus4482.0486.2487.0386.9787.8588.1088.053.95
Motus4582.2586.2286.9286.8587.6188.0287.993.97
Motus4682.3186.1987.0286.9187.7387.9488.023.86
MotusMean ± SD82.20 ± 0.1186.20 ± 0.0487.03 ± 0.0786.93 ± 0.0587.71 ± 0.0988.03 ± 0.0788.03 ± 0.063.96 ± 0.06
FastWAM4289.2889.7690.5691.0692.0091.9491.595.03
FastWAM4389.3289.5790.5491.0991.9791.8291.665.00
FastWAM4489.2489.7790.5590.9792.0691.7891.695.02
FastWAM4589.3889.8290.4391.1092.0891.9591.624.98
FastWAM4689.2089.6090.6591.0392.1991.8791.764.92
FastWAMMean ± SD89.28 ± 0.0789.70 ± 0.1190.55 ± 0.0891.05 ± 0.0592.06 ± 0.0991.87 ± 0.0791.66 ± 0.074.99 ± 0.04
LingBotVA4282.7188.9689.5790.3791.2892.1492.223.68
LingBotVA4382.7388.8689.6590.2791.5092.0192.273.67
LingBotVA4482.8488.8589.4590.4091.3792.0492.213.65
LingBotVA4582.8588.9889.5690.3491.3592.0892.313.57
LingBotVA4682.8389.1089.5590.4191.2592.1592.293.59
LingBotVAMean ± SD82.79 ± 0.0788.95 ± 0.1089.56 ± 0.0790.36 ± 0.0691.35 ± 0.1092.08 ± 0.0692.26 ± 0.043.63 ± 0.05
AuthorsOpenReview ↗

Response to W2 and Q1:Adaptive baselines and incremental gains from adaptive scheduling

Thank you for suggesting an adaptive-step baseline and evaluating the benefits of adaptive scheduling at matched actual latency. We have added the following results to Section 5.1 (Table 2) and Appendix I.

1. Base + adaptive. We freeze the undistilled model and train the same scheduler architecture to predict difficulty and budget-specific fidelity from first-pass features, using its own fine-grained trajectories as references and the same risk-conditioned selection rule. Compared with fixed four-step Base, SR improves by 0.62, 0.94, and 1.03 percentage points on Motus, FastWAM, and LingBotVA, respectively. Full AnyStep further improves SR by 3.44, 3.61, and 5.04 points over Base + adaptive.

BackboneMethodSR (%)Mean denoising steps
MotusBase, fixed 4 steps83.904
MotusBase + adaptive84.524.24
Motus (ours)AnyStep + adaptive (Ours)87.963.98
FastWAMBase, fixed 4 steps87.044
FastWAMBase + adaptive87.984.57
FastWAM (Ours)AnyStep + adaptive (Ours)91.595.03
LingBotVABase, fixed 4 steps86.154
LingBotVABase + adaptive87.183.79
LingBotVA (ours)AnyStep + adaptive (Ours)92.223.68

2. Incremental gains from adaptive scheduling. We measure latency on a single NVIDIA A100. At neighboring fixed budgets, AnyStep improves SR over Base by 3.14–3.76 points. To isolate scheduling, we compare the same student at fixed integer budgets whose latencies bracket adaptive inference. Adaptive SR exceeds both neighboring settings across all three backbones. Compared with the higher budgets (4, 6, and 4 steps for Motus, FastWAM, and LingBotVA), SR improves by 0.92, 0.53, and 2.65 points while latency decreases by 11.8%, 9.8%, and 7.1%, respectively. Compared with fixed ten-step students, adaptive SR remains within 0.35 points while latency decreases by 41.4%–56.2%.

BackboneStepsSR (%)Latency (ms/call)
Motus (base)383.42765
Motus (ours fixed budget)386.73793
Motus (ours adaptive budget)3.98 (Adaptive)87.96859
Motus (base)483.90969
Motus (ours fixed budget)487.04974
Motus (base)1087.841932
Motus (ours fixed budget)1088.091959
FastWAM (base)587.22271
FastWAM (ours fixed budget)590.98276
FastWAM (ours adaptive budget)5.03 (Adaptive)91.59295
FastWAM (base)687.74309
FastWAM (ours fixed budget)691.06327
FastWAM (base)1091.83496
FastWAM (ours fixed budget)1091.94503
LingBotVA (base)385.582914
LingBotVA (ours fixed budget)388.722986
LingBotVA (ours adaptive budget)3.68 (Adaptive)92.223247
LingBotVA (base)486.153403
LingBotVA (ours fixed budget)489.573495
LingBotVA (base)1090.395972
LingBotVA (ours fixed budget)1092.146033

3. More Analysis of incremental gains from adaptive scheduling. We further analyze Motus on all five tasks where the original Base Motus achieves SR below 70%, averaged across clean and randomized settings. Adaptive inference achieves 62.40% SR at 4.67 mean steps, improving over the fixed 4-step and 10-step AnyStep students by 7.00 and 5.40 percentage points, respectively. Holding the distilled student fixed allows us to assess the adaptive scheduler’s additional contribution separately from distillation, including a comparison at comparable budgets. Many RoboTwin tasks already achieve relatively high SR with both Base and AnyStep, so averaging across all tasks dilutes the larger gains observed on difficult manipulation tasks. We thank the reviewer for encouraging this analysis and included these task-level comparisons in the revised paper (Table 2).

TaskBase 2 StepAnyStep 2 StepBase 3 StepAnyStep 3 StepBase 4 StepAnyStep 4 StepBase 6 StepAnyStep 6 StepBase 10 StepAnyStep 10 StepAdaptive SR (%)Adaptive Steps
Blocks Ranking Size57.0057.5056.0057.5057.5058.0059.0059.5069.0066.0072.503.81
Hanging Mug24.0034.0024.5034.0025.0039.0030.5034.5038.0036.0044.506.45
Move Can Pot44.0045.5045.0053.0045.5053.5049.5054.5054.0050.5060.003.46
Place Mouse Pad52.5056.5052.5057.0053.0057.5059.5060.0067.0066.5066.503.84
Scan Object56.5065.0058.0064.0058.5069.0063.5067.5066.5066.0068.505.81
Average46.8051.7047.2053.1047.9055.4052.4055.2058.9057.0062.40(+7.00 pp vs AnyStep fixed 4 step, + 5.40 pp vs AnyStep fixed 10 step)4.67
AuthorsOpenReview ↗

Response to W3 and Q3: Preview reuse and step counting

W3 and Q3: Preview reuse and step counting

We apologize for the unclear description in Appendix D. In our implementation, AnyStep+Gate and our method both reuse the one-step preview, and the step counts reported for both methods in Table 1 already include this evaluation. Thus, Table 1 compares the two schedulers under the same preview-reuse procedure. We have corrected the descriptions in Table 1 and Appendix D (TRAINING-FREE GATING BASELINE) and explicitly clarified the step-counting convention.

BackboneMethodClean SR (%)Clean StepsRandom SR (%)Random StepsAverage SR (%)Average Steps
MotusAnyStep+Gate87.605.2386.924.9587.265.09
MotusAnyStep (Ours)88.124.0287.803.9487.963.98
FastWAMAnyStep+Gate90.286.3490.046.9890.166.66
FastWAMAnyStep (Ours)92.125.0291.065.0391.595.03
LingBotVAAnyStep+Gate90.164.5890.525.0490.344.81
LingBotVAAnyStep (Ours)92.743.5491.703.8292.223.68

Comments on the paper

2 comments

Dear Reviewers and Area Chairs,

Thank you very much for your careful reviews and constructive suggestions. Your feedback has helped us strengthen the experimental evidence and improve the clarity of our paper. We have revised the manuscript and provided point-by-point responses to each reviewer. The main changes are marked in blue in the revised paper:

1. Isolated the gains from adaptive scheduling (Section 5.1, Table 2, and Appendix I).
Addresses: Reviewer dhCG (W1, W2, Q1); Reviewer EZcz (W1, Q1).
We added Base + adaptive baselines and actual latency comparisons. On all five tasks where Base Motus achieves SR below 70%, adaptive inference improves SR over the same student at fixed four and ten steps by 7.00 and 5.40 points, respectively, using 4.67 mean steps.

2. Evaluated training-seed robustness (Appendix H).
Addresses: Reviewer dhCG (W1, Q1); Reviewer WuyW (W1, Q1).
We trained the student and scheduler with five seeds for all three backbones. The sample SD of adaptive SR is at most 0.07 percentage points.

3. Validated the step-selection signals (Appendices E.1–E.5).
Addresses: Reviewer dhCG (W5, W6, Q2); Reviewer EZcz (W2, Q2); Reviewer WuyW (W2, Q2).
Selected higher budgets improve paired-rollout SR from 81.8% to 87.9%. Scheduler ablations show 1.22/1.40-point clean/randomized SR gains over constant thresholds at comparable budgets. We also added task- and phase-level analyses of risk and selected budgets.

4. Expanded distillation comparisons (Appendix F.2).
Addresses: Reviewer 5Wza (Q1, Q2).
We evaluated LCM and DMD2 and added AnyFlow results. Adaptive AnyStep exceeds four-step LCM and DMD2 by 5.44 and 4.28 SR points, respectively, using 3.98 mean steps.

5. Extended benchmark and embodiment evaluation (Section 5 and Appendix J).
Addresses: Reviewer WuyW (S2).
We added single-arm Franka experiments on LIBERO and zero-shot transfer to LIBERO-Plus, achieving 98.2% SR at 4.45 mean steps and 50.8% SR at 5.44 mean steps, respectively.

6. Quantified offline costs (Appendices C.5–C.9).
Addresses: Reviewer dhCG (W4, Q4); Reviewer 5Wza (W1); Reviewer WuyW (Q5).
We report GPU-hours across training and trajectory-generation stages, distillation peak memory, and the costs of teacher trajectories and scheduler dataset construction.

7. Expanded sensitivity analyses and loss ablations (Appendices E.6, C.2, and E.3).
Addresses: Reviewer dhCG (W6); Reviewer WuyW (Q3, Q4).
We explain and evaluate ten threshold configurations, retrain the student and scheduler with two alternative loss-weight combinations, and ablate the standard flow-matching loss.

8. Improved presentation and clarified execution (Sections 1, 4, 5.1, and Appendix D).
Addresses: Reviewer dhCG (W3, Q3, W7); Reviewer 5Wza (Q2); Reviewer WuyW (S1).
We clarify the central contributions, discuss recent methods, and distinguish offline supervision from online inference. We also clarify that AnyStep+Gate and our scheduler share preview reuse and include its cost in their step counts.

Thank you again for your time and constructive feedback.

Best regards,
The Authors

Dear Reviewers, 

We’re about halfway through the discussion period. Please read the authors’ rebuttal and assess whether it adequately addresses the key concerns raised in your review. We encourage you to leave a follow-up comment for the authors, indicating which concerns have been resolved and whether any important issues remain or require further clarification. Please update your review and scores as appropriate in light of the rebuttal and discussion.

Thank you for your time and participation. 

Best regards, 

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.